PULSE

Resorts usually learn a guest is unhappy from a front-desk complaint or a bad review, when the stay is already ruined and the only fix is a refund. The warning signs exist earlier (no door activity for hours, a skipped meal, a late room-service order) but they sit in separate systems (door locks, POS, Wi-Fi, phone) that nobody watches together.

A live control centre that scores every guest's risk of a bad stay from those small signals and tells staff what to do before the guest complains.

Live alert queue

Guests are ranked by a 1–10 risk score and re-sorted as new signals arrive over a WebSocket, with no polling.

Risk evaluation panel

For the selected guest it shows why the alert fired, the raw signals behind it, similar past cases and how they ended, and a confidence figure.

Recommended actions

Each action is tagged by team (concierge, marketing, front desk) with a predicted success rate. One click resolves the alert and logs it on the guest's timeline.

Hybrid scoring

Fixed rules own the number: a service delay over 30 minutes adds 4, a direct complaint adds 5. Gemini only explains the risk and proposes actions, grounded in the hotel's SOP document and the three closest matches from 200 historical recovery cases.

Simulation engine

A background ticker generates guest activity every 10 seconds, with controls to escalate one guest, trigger a kitchen outage that cascades across rooms, or reset, so the whole system can be demoed without a real hotel.

Onboarding and guided tour

A first-run walk through the three panels for someone seeing it for the first time.

Solo developer. I designed the product, built the React front end and the FastAPI back end, wrote the rule engine, state engine, retrieval service and simulator, and designed the control-centre UI.

Front end

React 19, TypeScript, Vite 6, Tailwind CSS v4, Base UI / shadcn, Framer Motion, Recharts, Sonner

Back end

Python 3.12, FastAPI, Uvicorn, SQLModel

Data

SQLite in WAL mode, so simulator writes don't lock API reads

AI

Gemini 2.5 Flash in JSON mode; Gemini embeddings with NumPy cosine similarity for retrieval (no vector database)

Real time

Native WebSockets

Safety rails

SlowAPI rate limiting, and a rule-based fallback when the model call fails

Make AI failures visible

When the model call fails the panel shows fallback content that looks like a real analysis. It should say "rule-based only" so staff know what they are reading.

Connect real data sources

Everything runs on the simulator today. The ingestion endpoint exists; the next step is adapters for an actual PMS, POS and door-lock feed.

Cache embeddings

Precomputing and storing case embeddings would cut analysis latency and API cost.

Real tests

Proper assertions for the rule and state engines in place of a script that prints results.

A working rules editor

It would let managers tune thresholds without a developer.

Measure outcomes

The predicted success rates are illustrative. Tracking which actions actually prevented a complaint would let the system learn them.